Bidirectional LSTM + l2-normalization for image weights

[AndreiBarsan]

• Experiment logs are now being dumped (in a somewhat dirty way) to a file in the experiment folder.
• Option to l2-normalize pretrained VGG weights, as suggested by the VQA paper baselines (doesn't seem to help, but may help in other configurations).
• Bidirectional LSTM support via a flag.